Timeline for Indicator variables and backward elimination with GLM
Current License: CC BY-SA 3.0
11 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jul 28, 2017 at 13:20 | comment | added | Stephan Kolassa | @ttnphns: thanks for the pointer. I looked at the discussion and honestly can't make a decision - both sides have very good arguments. So I'll refrain from voting. | |
| Jul 28, 2017 at 9:47 | comment | added | ttnphns | Stephan, can you kindly visit the dilemma if we ought to create tag "dummy-variables" and leave there a comment either con or pro? | |
| Jan 18, 2013 at 12:29 | comment | added | Stephan Kolassa | It all depends on your data, of course. One example which I encountered in real life: retailers run promotions (like multi-page flyers in people's mailboxes) to increase sales, and a promotion predictor may have levels like "ad on the first page of a flyer", "ad on the second page", "ad on the last page" and so on. However, it is unclear whether the flyer page actually makes a difference, so collapsing all levels into one level or into two levels "ad on first page" and "ad somewhere else" may be better in terms of parsimony and predictive power. | |
| Jan 18, 2013 at 12:21 | comment | added | sim | come to think of it, i would have one last question if you don't mind:i think i get your point about data dredging, since the enumerative approach (testing all subsets & combinations) is an exhaustive in-sample optimization of which the result is imposed on new instances (for prediction). am i correct so far? in your example, quite clear that the reduction of dfs will reduce the AIC - accompanied by the knowledge about, as you said, how the data was generated. however, since i seldom know about that in practice, could you give an example when you would collapse factors? thanks again! | |
| Jan 18, 2013 at 12:05 | history | edited | Stephan Kolassa | CC BY-SA 3.0 |
corrected glm family argument from poisson to binomial
|
| Jan 18, 2013 at 11:44 | comment | added | sim | thanks a lot for the additional comments & the illustrative example! it's much clearer to me now. | |
| Jan 18, 2013 at 11:02 | comment | added | Stephan Kolassa | I edited my answer; clearer now? | |
| Jan 18, 2013 at 11:01 | history | edited | Stephan Kolassa | CC BY-SA 3.0 |
added clarification
|
| Jan 18, 2013 at 10:21 | vote | accept | sim | ||
| Jan 18, 2013 at 10:21 | comment | added | sim | Thanks for your reply, can't vote up or accept yet, but will do as soon as I can. Could you please elaborate on the collapsing part? Say I had 4 levels, would I simply combine a subset of levels to an artificial new one whilst keeping the others (subsequently enumerating through all proper subsets and subset-combinations of the indicator)? | |
| Jan 18, 2013 at 10:08 | history | answered | Stephan Kolassa | CC BY-SA 3.0 |